Composition-Aware Image Steganography through Adversarial Self-Generated Supervision

Abstract

Steganography is an important and prevailing information hiding tool to perform secret message transmission in an open environment. Existing steganography methods can mainly fall into two categories: pre-defined rule-based and data-driven methods. The former is susceptible to the statistical attack while the latter adopts the deep convolution neural networks to promote security under statistical attack. However, the deep learning-based methods suffer from perceptible artificial artifacts. In this paper, we introduce a novel Composition-Aware Image Steganography termed \textbf{CAIS} to guarantee both visual security and robustness to attack through self-generated supervision. The key innovation is an adversarial composition estimation module to integrate rule-based and deep generative adversarial methods. We perform a rule-based image blending method to obtain infinite synthetically data-label pairs and perform an auxiliary adversarial composition estimation task. The innovative self-generated supervision could largely promote the ability to recognize message patterns from steganographic outputs, which results in better steganography performance. Furthermore, an effective Global-and-Part checking is designed to alleviate visual artifacts caused by hiding secret information. We conduct a comprehensive analysis of CAIS from various aspects such as security and robustness to verify the superior performance of the proposal. Experimental results on three large-scale widely-used datasets show the superior performance of our CAIS compared with several state-of-the-art approaches.